Communication efficient sparse-reduce in distributed machine learning
View Patent ↗A method is provided for sparse communication in a parallel machine learning environment. The method includes determining a fixed communication cost for a sparse graph to be computed. The sparse graph is (i) determined from a communication graph that includes all the machines in a target cluster of the environment, and (ii) represents a communication network for the target cluster having (a) an overall spectral gap greater than or equal to a minimum threshold, and (b) certain information dispersal properties such that an intermediate output from a given node disperses to all other nodes of the sparse graph in lowest number of time steps given other possible node connections. The method further includes computing the sparse graph, based on the communication graph and the fixed communication cost. The method also includes initiating a propagation of the intermediate output in the parallel machine learning environment using a topology of the sparse graph.
1. A computer-implemented method for sparse communication in a parallel machine learning environment, comprising:
determining, by a processor, a fixed communication cost for a sparse graph to be computed, wherein the sparse graph is (i) determined from a communication graph that includes all the machines in a target cluster of the parallel machine learning environment, and (ii) represents a communication network for the target cluster having (a) an overall spectral gap greater than or equal to a minimum threshold, and (b) certain information dispersal properties such that an intermediate output from a given node disperses to all other nodes of the sparse graph in lowest number of time steps given other node connections;
computing, by the processor, the sparse graph, based on the communication graph and the fixed communication cost; and
initiating, by the processor, a propagation of the intermediate output in the parallel machine learning environment using a topology of the sparse graph,
wherein the method further comprises automatically selecting a number and a type of model replicas for computing the sparse graph, based on the overall spectral gap and a number of edges in the communication graph.
2. The computer-implemented method of claim 1 , wherein the fixed communication cost is determined empirically.
3. The computer-implemented method of claim 1 , wherein the fixed communication cost is determined to allow selection of an out-degree for each node in the sparse graph.
4. The computer-implemented method of claim 1 , wherein the intermediate output is a model update for model training.
5. The computer-implemented method of claim 1 , wherein the parallel machine learning environment is a distributed parallel machine learning environment.
6. The computer-implemented method of claim 1 , wherein the parallel machine learning environment is comprised in a surveillance system.
7. The computer-implemented method of claim 1 , wherein the parallel machine learning environment is comprised in a language translation system.
8. The computer-implemented method of claim 1 , wherein the parallel machine learning environment is comprised in an image recognition system.
9. The computer-implemented method of claim 1 , wherein the method is performed to train a parallel model in a distributed manner using multiple computing nodes of the target cluster.
10. A computer program product for sparse communication in a parallel machine learning environment, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:
determining, by a processor of the computer, a fixed communication cost for a sparse graph to be computed, wherein the sparse graph is (i) determined from a communication graph that includes all the machines in a target cluster of the parallel machine learning environment, and (ii) represents a communication network for the target cluster having (a) an overall spectral gap greater than or equal to a minimum threshold, and (b) certain information dispersal properties such that an intermediate output from a given node disperses to all other nodes of the sparse graph in lowest number of time steps given other node connections;
computing, by the processor, the sparse graph, based on the communication graph and the fixed communication cost; and
initiating, by the processor, a propagation of the intermediate output in the parallel machine learning environment using a topology of the sparse graph,
wherein the method further comprises automatically selecting a number and a type of model replicas for computing the sparse graph, based on the overall spectral gap and a number of edges in the communication graph.
11. The computer program product of claim 10 , wherein the fixed communication cost is determined empirically.
12. The computer program product of claim 10 , wherein the fixed communication cost is determined to allow selection of an out-degree for each node in the sparse graph.
13. The computer program product of claim 10 , wherein the intermediate output is a model update for model training.
14. The computer program product of claim 10 , wherein the parallel machine learning environment is a distributed parallel machine learning environment.
15. The computer program product of claim 10 , wherein the parallel machine learning environment is comprised in a surveillance system.
16. The computer program product of claim 10 , wherein the parallel machine learning environment is comprised in a language translation system.
17. The computer program product of claim 10 , wherein the parallel machine learning environment is comprised in an image recognition system.
18. A computer processing system for sparse communication in a parallel machine learning environment, comprising:
a memory for storing program code; and
a processor for running the program code to:
determine a fixed communication cost for a sparse graph to be computed, wherein the sparse graph is (i) determined from a communication graph that includes all the machines in a target cluster of the parallel machine learning environment, and (ii) represents a communication network for the target cluster having (a) an overall spectral gap greater than or equal to a minimum threshold, and (b) certain information dispersal properties such that an intermediate output from a given node disperses to all other nodes of the sparse graph in lowest number of time steps given other node connections;
compute the sparse graph, based on the communication graph and the fixed communication cost; and
initiate a propagation of the intermediate output in the parallel machine learning environment using a topology of the sparse graph,
wherein processor further runs the program code to automatically select a number and a type of model replicas for computing the sparse graph, based on the overall spectral gap and a number of edges in the communication graph.